systems
331
similarity of real world cases, which are fuzzy, that is, continuous and not discrete. Hansen, 2000.
There are at least four advantages of using fuzzy techniques in the retrieval stage Jeng and Liang, 1995. First, it allows numerical features to be converted into
fuzzy terms to simplify comparison. For example we can convert the age of a patient into a categorical scale e.g., old, middle-aged, or young. Second, fuzzy sets allow
multiple indexing of a case on a single feature with different degrees of membership. This increases the flexibility of case matching. For example, a 50-year old patient may
be classified as old 0.6 and middle-aged 0.5 where 0.6 and 0.5 are the degrees that the 50-year old is classified as old and middle-aged respectively. This allows the case
to be viewed as a candidate when we are looking for either an old patient or a middle- aged patient. Third, fuzzy sets make it easier to transfer knowledge across domains.
For instance, we have cases showing persons older than 50 years of age i.e. old persons will need special effort to get a good job. We can use these cases to derive a
guideline that computer software older than 2 years on the market i.e. old software will need special effort to make a profit. The absolute age scales are different in these
two domains but the fuzzy transformation provides a bridge for comparison. Finally, fuzzy sets allow term modifiers to be used to increase the flexibility in case retrieval.
For example we can search very old patients from a case base containing old patients with possibilities ranging from 0.5 to 1.0. Here “very” is a modifier of “old”, which
can be used to modify the membership grade of old and result in a subset of old patients considered very old being retrieved. This enhances the flexibility of
retrieval.
4. THE F-CIR SYSTEM: FUZZY CASE IDENTIFICATION
In this paper fuzzy logic is applied in case based reasoning systems. The system produced is called F-CIR Fuzzy Case Identifier and Retriever The case based
reasoning system must be able to identify cases ready for retrieval. If the system is not able to properly identify a suitable case this case may not be retrieved, although it
might be useful. Case adaptation may be needed for the retrieval stage to work efficiently. Case attributes can be either qualitative or quantitative. Qualitative
attributes have discrete nominal values.
Case base reasoners usually work with nominal categories and problems most often occur most often occur with continuous attributes. It is necessary to devise a way
to translate quantitative values to nominal ones. If we try to classify cases using classic “crisp” sets we will probably encounter problems: Classic sets do not describe
qualitative attributes adequately because they give the same degree of membership to all their members and the same degree of exclusion to all the non-members. Let’s
discuss for example a system that is called to decide upon the appropriate medication for a hospital patient. One factor on which the choice of the appropriate medication
depends is the patient’s age. If we have a “crisp” set classification procedure then we
systems
332
have to set arbitrary limits to say when a man is old. This kind of classification may have its benefits but it also has its drawbacks: If we say that this limit is at 60 years of
age, then for our system a 61 year old man is as “old” as a 90 year old, but in reality they are very different in their degree of oldness. A man aged 61 is “old” but a man of
90 is “definitely old”. Another problem with the traditional approach is that it does not provide adequate flexibility for marginal cases. A man aged 59 years and eleven
months is classifies as a “not old” when a man aged 60 is classified as “old”, even thought they practically have the same age. This is not the way humans see things and
can lead to problems. If, for example, we consider medical care in a hospital for a patient near the age limit, classifying him in one category can result in treatment not
suitable for him. With fuzzy set membership functions we can have gradual membership to a set. Case based reasoners using the standard set theory are risking
taking the wrong decisions by making wrong assumptions. Misinterpretations of a given situation can lead to errors. Suppose we have a system which searches for
people who are “young and rich” and let’s say that someone is defined as “young” if his age is less than 30 and “rich” if his wealth is more than 1.000.000. Suppose now
that we have three persons whose age and wealth are 26, 50.000, 67, 10.000.000 and 30, 999.990. None of them qualifies the criteria as being “young and rich”.
The first is young but not rich, the second rich but not young, and the third is neither. A classic case based reasoner may retrieve the first or the second person based on
partial match but it will miss the third, although in reality he is the one closest to the criteria. Using fuzzy sets can overcome the above problems. A person may have
different degrees of membership in sets that would be mutually exclusive with the classic set membership definition. He can be at the same time classified as young and
old, with different degrees of membership.
A two stage process is applied in F-CIR. The first stage is training the system through the fuzzification of the cases. The second is the traversal of the fuzzyfied case
base to find the best matching case. The fuzzification works both automatically and also biased by expert tuning. Expert perceptiveness is utilized to eliminate the risk of
misinterpretation of the situation. The attributes can be either discriminative or continuous. The fuzzification process is used for the quantitative, continuous
attributes. The fuzzyfier describes the way a continuous attribute will be turned into a classifiable one.
The fuzzification process works according the following steps: 1.
Eliminate one attribute and test try to find irrelevant attributes.
2. Eliminate all but one attributes and test find the relative significance of
